Automated Volumetric Segmentation of Microaneurysms on OCT Using Artificial Intelligence.

Purpose: To develop and validate a deep learning-based method for the automated identification and volumetric segmentation of microaneurysms (MAs) in diabetic retinopathy (DR) using optical coherence tomography (OCT). Design: Cross-sectional study with a representative longitudinal case analysis. Participants: A total of 125 participants were enrolled, including 20 healthy eyes, 27 with mild nonproliferative DR (NPDR), 30 with moderate NPDR, 30 with severe NPDR, and 18 with proliferative DR. Methods: We obtained multiple repeated 3×3-mm scans from each participant using a commercial 120-kHz spectral-domain OCT system (Solix; Visionix/Optovue, Inc., California, USA) which were registered and averaged to generate high-definition volumes. We developed an end-to-end three-dimensional convolutional neural network, incorporating an encoder-decoder architecture with residual blocks, to segment MAs volumetrically. The input to the network consists of the original OCT volume concatenated with its reflectance-inverted counterpart to enhance feature extraction. Expert graders manually delineated MAs to generate initial annotations, which were later refined into a verified ground truth through an iterative human-in-the-loop refinement process to capture MAs initially overlooked or mischaracterized. Both single-scan volumes and volumes averaged across 4-10 repeated OCT volumes were used to train and test the model. We evaluated model performance at multiple levels, including voxel-level segmentation, lesion-level detection, and eye-level diagnosis. We also analyzed the differences in MA counts across various stages of DR severity and quantified the MA number and volume distribution in different retinal layers. A longitudinal analysis of MA dynamics was performed in a case with severe NPDR and macular edema using the algorithm. Main Outcome Measures: Voxel-level MA segmentation accuracy, MA quantification categorized by retinopathy severity, and longitudinal analysis of MA dynamics in OCT volumes. Results: In the test dataset (20 healthy, 20 DR eyes), when evaluated against the verified ground truth, the algorithm demonstrated high voxel-level accuracy, with F1 scores of 79.2% on single volumes and 86.1% on averaged volumes. Lesion-level detection reached an F1 score of 96.0% (single) and 97.1% (averaged), while scan-level diagnostic accuracy for MA presence was 97.5% (single and averaged). MA counts significantly increased from mild to moderate NPDR (7.4±1.5 vs. 12.7±1.6, P = 0.014) but not between any other severity comparisons (P>0.05). MAs were primarily localized to the inner nuclear layer and outer plexiform layer. In one case with severe NPDR and macular edema, longitudinal analysis tracked 11 resolved MAs and 3 new MAs post-treatment, corresponding to a reduction in fluid volume. Conclusions: A deep learning-based method can accurately identify and segment MAs volumetrically on OCT, enabling quantification and characterization and potentially aiding in the diagnosis, monitoring, and management of DR.